/// can be deleted.
bool IsLoaded = false;
+ /// Number of stores to the global.
+ unsigned NumStores = 0;
+
/// Keep track of what stores to the global look like.
enum StoredType {
/// There is no store to this global. It can thus be marked constant.
}
}
+// For a global variable with one store, if the store dominates any loads,
+// those loads will always load the stored value (as opposed to the
+// initializer), even in the presence of recursion.
+static bool forwardStoredOnceStore(
+ GlobalVariable *GV, const StoreInst *StoredOnceStore,
+ function_ref<DominatorTree &(Function &)> LookupDomTree) {
+ const Value *StoredOnceValue = StoredOnceStore->getValueOperand();
+ // We can do this optimization for non-constants in nosync + norecurse
+ // functions, but globals used in exactly one norecurse functions are already
+ // promoted to an alloca.
+ if (!isa<Constant>(StoredOnceValue))
+ return false;
+ const Function *F = StoredOnceStore->getFunction();
+ SmallVector<LoadInst *> Loads;
+ for (User *U : GV->users()) {
+ if (auto *LI = dyn_cast<LoadInst>(U)) {
+ if (LI->getFunction() == F &&
+ LI->getType() == StoredOnceValue->getType() && LI->isSimple())
+ Loads.push_back(LI);
+ }
+ }
+ // Only compute DT if we have any loads to examine.
+ bool MadeChange = false;
+ if (!Loads.empty()) {
+ auto &DT = LookupDomTree(*const_cast<Function *>(F));
+ for (auto *LI : Loads) {
+ if (DT.dominates(StoredOnceStore, LI)) {
+ LI->replaceAllUsesWith(const_cast<Value *>(StoredOnceValue));
+ LI->eraseFromParent();
+ MadeChange = true;
+ }
+ }
+ }
+ return MadeChange;
+}
+
/// Analyze the specified global variable and optimize
/// it if possible. If we make a change, return true.
static bool
if (optimizeOnceStoredGlobal(GV, StoredOnceValue, DL, GetTLI))
return true;
+ // Try to forward the store to any loads. If we have more than one store, we
+ // may have a store of the initializer between StoredOnceStore and a load.
+ if (GS.NumStores == 1)
+ if (forwardStoredOnceStore(GV, GS.StoredOnceStore, LookupDomTree))
+ return true;
+
// Otherwise, if the global was not a boolean, we can shrink it to be a
// boolean. Skip this optimization for AS that doesn't allow an initializer.
if (SOVConstant && GS.Ordering == AtomicOrdering::NotAtomic &&
if (SI->isVolatile())
return true;
+ ++GS.NumStores;
+
GS.Ordering = strongerOrdering(GS.Ordering, SI->getOrdering());
// If this is a direct store to the global (i.e., the global is a scalar
@foo = internal global i32 0, align 4
-define dso_local i32 @bar() {
+define void @store() {
entry:
store i32 5, i32* @foo, align 4
+ ret void
+}
+
+define i32 @bar() {
+entry:
%0 = load i32, i32* @foo, align 4
ret i32 %0
}
;CHECK: @bar
;CHECK-NEXT: entry:
-;CHECK-NEXT: store i1 true, i1* @foo, align 1, !dbg ![[DbgLocStore:[0-9]+]]
;CHECK-NEXT: %.b = load i1, i1* @foo, align 1, !dbg ![[DbgLocLoadSel:[0-9]+]]
;CHECK-NEXT: %0 = select i1 %.b, i32 5, i32 0, !dbg ![[DbgLocLoadSel]]
;CHECK-NEXT: call void @llvm.dbg.value({{.*}}), !dbg ![[DbgLocLoadSel]]
;CHECK-NEXT: ret i32 %0, !dbg ![[DbgLocRet:[0-9]+]]
-;CHECK: ![[DbgLocStore]] = !DILocation(line: 1,
-;CHECK: ![[DbgLocLoadSel]] = !DILocation(line: 2,
-;CHECK: ![[DbgLocRet]] = !DILocation(line: 3,
+;CHECK: ![[DbgLocLoadSel]] = !DILocation(line: 3,
+;CHECK: ![[DbgLocRet]] = !DILocation(line: 4,
; Negative test for AS(3). Skip shrink global to bool optimization.
; CHECK: @lvar = internal unnamed_addr addrspace(3) global i32 undef
-define void @test_global_var() {
+define void @test_global_var(i1 %i) {
; CHECK-LABEL: @test_global_var(
; CHECK: store volatile i32 10, i32* undef, align 4
;
entry:
+ br i1 %i, label %bb1, label %exit
+bb1:
store i32 10, i32* @gvar
br label %exit
exit:
ret void
}
-define void @test_lds_var() {
+define void @test_lds_var(i1 %i) {
; CHECK-LABEL: @test_lds_var(
; CHECK: store i32 10, i32 addrspace(3)* @lvar, align 4
; CHECK: [[LD:%.*]] = load i32, i32 addrspace(3)* @lvar, align 4
; CHECK: store volatile i32 [[LD]], i32* undef, align 4
;
entry:
+ br i1 %i, label %bb1, label %exit
+bb1:
store i32 10, i32 addrspace(3)* @lvar
br label %exit
exit:
define i1 @dom_const() {
; CHECK-LABEL: @dom_const(
-; CHECK-NEXT: store i1 true, ptr @g1, align 1
; CHECK-NEXT: call void @b()
-; CHECK-NEXT: [[R:%.*]] = load i1, ptr @g1, align 1
-; CHECK-NEXT: ret i1 [[R]]
+; CHECK-NEXT: ret i1 true
;
store i1 true, ptr @g1
call void @b()
; CHECK-LABEL: @dom_multiple_function_loads(
; CHECK-NEXT: store i1 true, ptr @g6, align 1
; CHECK-NEXT: call void @b()
-; CHECK-NEXT: [[R:%.*]] = load i1, ptr @g6, align 1
-; CHECK-NEXT: ret i1 [[R]]
+; CHECK-NEXT: ret i1 true
;
store i1 true, ptr @g6
call void @b()
define i32 @main() {
; CHECK-LABEL: @main(
; CHECK-NEXT: entry:
-; CHECK-NEXT: store i1 true, i1* @a, align 4
; CHECK-NEXT: [[TMP0:%.*]] = load i32*, i32** @e, align 8
; CHECK-NEXT: store i32 0, i32* [[TMP0]], align 4
; CHECK-NEXT: store i32* null, i32** @e, align 8